Expose Camouflage in the Water: Underwater Camouflaged Instance Segmentation and Dataset

Area of Science:

• Computer Vision
• Marine Biology
• Image Processing

Background:

• Underwater environments present significant challenges for computer vision tasks due to color distortion, low contrast, and blurring.
• Existing camouflaged instance segmentation (CIS) methods struggle with underwater scenes due to limited training data and environmental degradation.

Purpose of the Study:

• To introduce the first dedicated dataset for underwater camouflaged instance segmentation (UCIS4K).
• To propose a novel network (UCIS-SAM) for enhanced underwater camouflaged instance segmentation.

Main Methods:

• Developed the UCIS4K dataset with 3,953 annotated images of camouflaged marine organisms.
• Proposed the Underwater Camouflaged Instance Segmentation network based on Segment Anything Model (UCIS-SAM).
• UCIS-SAM incorporates Channel Balance Optimization Module (CBOM), Frequency Domain True Integration Module (FDTIM), and Multi-scale Feature Frequency Aggregation Module (MFFAM).

Main Results:

• The proposed UCIS-SAM network demonstrated superior performance compared to state-of-the-art methods on the UCIS4K dataset and public benchmarks.
• The CBOM module improved underwater feature learning, FDTIM emphasized object features, and MFFAM enhanced boundary segmentation for low-contrast instances.

Conclusions:

• The UCIS4K dataset and UCIS-SAM network represent significant advancements in underwater camouflaged instance segmentation.
• The proposed methods effectively address the challenges of segmenting camouflaged marine life in degraded underwater environments.